pyvene-interventions
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFE
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the
pyvenepackage viapipand downloads models or interventions from HuggingFace Hub (e.g.,meta-llama/Llama-2-7b-hf,zhengxuanzenwu/intervenable_honest_llama2_chat_7B). These are standard operations for the library using well-known services. - [COMMAND_EXECUTION]: Provides standard installation instructions (
pip install pyvene) and standard Python code for performing neural network interventions, training, and saving results. - [INDIRECT_PROMPT_INJECTION]: Involves processing user-supplied prompts for model interpretability tasks. The skill follows research methodologies with clear ingestion points (text prompts), boundary markers (string delimiters), and relies on standard transformers tokenization for input handling. Capabilities are limited to model analysis and saving intervention states.
Audit Metadata